Radar Human Activity Recognition with an Attention-Based Deep Learning Network.

Area of Science:

• Radar signal processing
• Machine learning
• Human-computer interaction

Background:

• Radar-based human activity recognition (HAR) offers non-contact sensing for applications like smart security and surveillance.
• Deep learning (DL) models show promise for HAR using micro-Doppler signals, but complex structures hinder real-time embedded use.
• Existing methods like traditional moving target indicator (MTI) have limitations in clutter suppression and accuracy.

Purpose of the Study:

• To propose an efficient deep learning network with an attention mechanism for radar-based HAR.
• To improve the computational efficiency and accuracy of HAR for real-time embedded systems.
• To enhance clutter suppression using an averaged cancellation method.

Main Methods:

• A novel network decouples Doppler and temporal features from radar signals.
• One-dimensional convolutional neural networks (1D CNN) extract Doppler features sequentially.
• Attention-mechanism-based long short-term memory (LSTM) processes Doppler features as time sequences.
• Averaged cancellation method enhances activity features and improves clutter suppression.

Main Results:

• The proposed method achieved an accuracy close to 96.9% on two human activity datasets.
• Recognition accuracy improved by approximately 3.7% compared to traditional MTI.
• The network structure is significantly more lightweight than comparable algorithms.
• Demonstrated superiority in expressiveness and computational efficiency over traditional methods.

Conclusions:

• The proposed efficient network with an attention mechanism is effective for radar-based HAR.
• The method offers improved accuracy and computational efficiency, suitable for real-time embedded applications.
• This approach shows great potential for advancing non-contact human activity recognition systems.